/ ˈɑː.kɑːʃ · bɪldz /
A person who builds production-grade AI systems and automations to eliminate repetitive work.
A mechanical engineer who kept automating his own job until it became the job.
aakash@builds:~$
I trained as a mechanical engineer.
At Eicher Tractors, I watched competent people spend entire days copying numbers between forms.
So I rebuilt the forms.
Crude, but the day got shorter.
At Idea Clan, I saw the same thing again: an editor cutting the same ad, one hook at a time.
So I built Magica to do the repetition.
Find the thing being done twice.
Delete the second time.
Now I build agents, voice systems and internal tools that remove repetitive work.
The art is the human.
The process is the machine.
Machines handle the repetition.
Humans handle the judgement.
That's what I build for.
> CURRENTLY
> TRACE
MODELS & REASONING
AUTOMATION
VOICE
SHIP IT
Scanned court filings — thousands of pages, no text layer, and material too sensitive to send to a hosted model.
A local-only pipeline: OCR, layout-aware chunking, embeddings and retrieval that never leave the machine, with citations back to the page image.
10,062 pages of scanned petitions turned into structured text — 2.8× what the PDFs' own text layer carried.
Invoices were being checked by eye. The ones that need catching are exactly the ones that look fine — arithmetic that doesn't add up, a total that doesn't match its own line items, the same invoice submitted twice.
The model is allowed to read and forbidden to compute. Gemini extracts what is printed on the page; JavaScript recomputes every number and disagrees out loud when the invoice doesn't add up. Send a PDF to a Telegram bot, get back PASS or a flag with reasons.
03
IN USE · Skyeverse · 2026
Scraped lead lists are mostly noise, and the judgement that separates a prospect from a name — is this person actually looking for clients — is the part a filter can't do.
Two scores, deliberately different in kind. A deterministic 100-point pass reads the facts on a profile: title, following, whether the bio talks about acquiring clients. A model then reads the same profile for buying readiness against a fixed schema. Final score is 40% the arithmetic, 60% the judgement; anything at or above 60 becomes a Notion record with an outreach hook attached.
Tell me the boring version of it — the spreadsheet, the inbox, the phone that keeps ringing. That's usually where the good automation is hiding.
akash.skyeverse@gmail.com